Motor Imagery EEG Signal Recognition Method Based On Depthwise Separable Convolution

With the advent of the intelligent era,brain-computer interfaces have made great progress in many fields.The core technology of brain-computer interfaces is to classify EEG signals.Deep learning technology provides a new method for the identification of EEG signal categories.Therefore,it is of great significance to carry out research based on deep learning to improve the accuracy of motor imagery EEG signal classification.The main research content and phased results of this project are as follows:First,in order to improve the signal-to-noise ratio of the motor imagery EEG signal,the data set in this article is processed by using ICA independent component analysis to remove artifacts and FIR bandpass filter;then based on the motor imagery EEG signal has rich time domain and frequency domain characteristics,this paper proposes a combined feature representation method(called feature time-frequency map),which is obtained by using a specific wavelet base to perform wavelet transformation on the EEG signal of a specific channel and time period,in order to obtain the best feature time-frequency map,this paper first uses the traditional machine learning method SVM to conduct a single study on the wavelet basis,signal period and channel to obtain the local optimum,and then combine them to obtain the global optimum;finally,the classification of motion imagination based on the optimal feature time-frequency map is carried out Performance research includes the following three aspects.(1)The research is based on the classification performance of motor imagery EEG signal under general convolutional neural network.The subject constructed a shallow CNN model and optimized the parameters through experiments.The experiment showed that the classification accuracy of EEG signals based on general convolutional neural networks is better than that based on traditional machine learning methods.(2)The research is based on the classification performance of the motor imagery EEG signal under the depthwise separable convolutional neural network.This paper proposes to use the depthwise separable convolutional neural network combined with the characteristic time-frequency map to perform the motor imagery EEG signal classification performance.The experimental results show that,this classification method can get a good classification effect.(3)Research on the classification performance of motor imagery EEG signals based on the combination of depthwise separable convolutional neural network and optimal feature time-frequency map.The classification accuracy rate using this combination reached 92.86%.The experimental results show that the combination of the depthwise separable convolutional neural network and the optimal feature time-frequency map has the best classification effect,which proves that the feature time-frequency map selected in this paper is the best choice,and the classification effect is better than many other current methods.The feasibility and effectiveness of the proposed method for motor imaging brain electrical signal classification are verified.